A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 13, 15:45 UTC
[Feature] Default `ignore_eos` True for `random` dataset (#28227)
Signed-off-by: yewentao256 <zhyanwentao@126.com>
Wentao Ye committed
Nov 7, 2025 at 07:35 UTC
4b1ff1322113aaf337573db90490f5741e65cee6
1 file changed
+8
vllm/benchmarks/serve.py
+8
@@ -1359,6 +1359,14 @@ async def main_async(args: argparse.Namespace) -> dict[str, Any]:
1359
"'--dataset-path' if required."
1360
)
1361
1362
+ # when using random datasets, default to ignoring EOS
1363
+ # so generation runs to the requested length
1364
+ if (
1365
+ args.dataset_name in ("random", "random-mm")
1366
+ and args.backend in OPENAI_COMPATIBLE_BACKENDS
1367
+ ):
1368
+ args.ignore_eos = True
1369
+
1370
# Load the dataset.
1371
input_requests = get_samples(args, tokenizer)
1372
goodput_config_dict = check_goodput_args(args)